Probabilistic approach to the dynamic ensemble selection using measures of competence and diversity of base classifiers

Abstract

In the paper measures of classifier competence and diversity using a probabilistic model are proposed. The multiple classifier system (MCS) based on dynamic ensemble selection scheme was constructed using both measures developed. The performance of proposed MCS was compared against three multiple classifier systems using six databases taken from the UCI Machine Learning Repository and the StatLib statistical dataset. The experimental results clearly show the effectiveness of the proposed dynamic selection methods regardless of the ensemble type used (homogeneous or heterogeneous). © 2011 Springer-Verlag.